Balancing Conservation and Development Through Explainable Machine Learning and NSGA-II: A Case Study of Osmaniye

Abstract

Land-use planning in ecologically sensitive landscapes requires balancing biodiversity conservation, ecosystem service provision, agricultural production, settlement expansion, and infrastructure demand within a single spatial system. This challenge is particularly significant in Mediterranean environments, where long-term land transformations and increasing development pressures intensify conflicts among competing land-use priorities. Accordingly, the present study develops an integrated spatial zoning and decision-support framework for Osmaniye Province, southern Türkiye. The framework integrates fuzzy multi-criteria evaluation, CatBoost-based machine learning, SHAP-based interpretability, and NSGA-II multi-objective optimization. The workflow followed a sequential decision process in which an expert-derived zoning surface was first established through fuzzy evaluation, reconstructed from continuous spatial predictors using CatBoost, interpreted through SHAP, and refined through NSGA-II under explicit spatial constraints. By using the expert-derived zoning surface as the learning target, the CatBoost stage aimed to evaluate the internal consistency and spatial learnability of the planning logic within a present-day zoning context. The results indicated that the integrated framework distinguished conservation, controlled-use, and development priorities while identifying the key environmental and anthropogenic drivers shaping class-specific zoning outcomes. The final zoning structure allocated 37.9% of the study area to conservation, 43.6% to controlled use, and 18.5% to development. The study shows that by including a transitional zone with varying proportions of conservation, controlled use, and development, a more balanced distribution among the three goals can be achieved compared to a fixed partition into these three zones. The findings further demonstrate that this approach is more effective than current zoning, which does not accommodate such trade-offs.